def nop_jump_junk(co, opc):
# Get the first jump target.
first_jump_inst = next(get_instructions_bytes(co.co_code, opc))
first_jump_target = first_jump_inst.argval
# Abuse the text header to find the second jump.
header_start_target = co.co_code.find(b'\r\n\r\n========')
has_pyprotect_header = header_start_target != -1
# Get the start of the JUMP_FORWARD, right before the header.
jump_forward_inst_size = instruction_size(opc.opmap['JUMP_FORWARD'], opc)
second_jump_start = header_start_target - jump_forward_inst_size
# Go over the existing code and decide whether to copy each byte.
fixed_code = bytearray()
for i in range(len(co.co_code)):
if (i < first_jump_target):
# Replace the first anti-disassembly jump with NOPs
fixed_code.append(opc.opmap['NOP'])
elif has_pyprotect_header and (i >= second_jump_start):
# Don't copy the second AD jump or anything after it.
pass
else:
# Copy everything else.
fixed_code.append(co.co_code[i])
# Set fixed code back on co.
co.co_code = fixed_code
co.freeze()
def op_range(self, start, end):
"""
Iterate through positions of opcodes, skipping
arguments.
"""
while start < end:
yield start
start += instruction_size(self.code[start], self.opc)
def get_inst(self, offset):
# Instructions can get moved as a result of EXTENDED_ARGS removal.
# So if "offset" is not in self.offset2inst_index, then
# we assume that it was an instruction moved back.
# We check that assumption though by looking at
# self.code's opcode.
if offset not in self.offset2inst_index:
offset -= instruction_size(self.opc.EXTENDED_ARG, self.opc)
assert self.code[offset] == self.opc.EXTENDED_ARG
return self.insts[self.offset2inst_index[offset]]
def build_prev_op(self):
"""
Compose 'list-map' which allows to jump to previous
op, given offset of current op as index.
"""
code = self.code
codelen = len(code)
# 2.x uses prev 3.x uses prev_op. Sigh
# Until we get this sorted out.
self.prev = self.prev_op = [0]
for offset in self.op_range(0, codelen):
op = code[offset]
for _ in range(instruction_size(op, self.opc)):
self.prev_op.append(offset)
def detect_control_flow(self, offset, targets, extended_arg):
"""
Detect structures and their boundaries to fix optimized jumps
in python2.3+
"""
# TODO: check the struct boundaries more precisely -Dan
code = self.code
op = code[offset]
# Detect parent structure
parent = self.structs[0]
start = parent['start']
end = parent['end']
# Pick inner-most parent for our offset
for struct in self.structs:
current_start = struct['start']
current_end = struct['end']
if ((current_start <= offset < current_end)
and (current_start >= start and current_end <= end)):
start = current_start
end = current_end
parent = struct
if op == self.opc.SETUP_LOOP:
# We categorize loop types: 'for', 'while', 'while 1' with
# possibly suffixes '-loop' and '-else'
# Try to find the jump_back instruction of the loop.
# It could be a return instruction.
start += instruction_size(op, self.opc)
target = self.get_target(offset, extended_arg)
end = self.restrict_to_parent(target, parent)
self.setup_loops[target] = offset
if target != end:
self.fixed_jumps[offset] = end
(line_no, next_line_byte) = self.lines[offset]
jump_back = self.last_instr(start, end, self.opc.JUMP_ABSOLUTE,
next_line_byte, False)
if jump_back:
jump_forward_offset = xdis.next_offset(code[jump_back], self.opc, jump_back)
else:
jump_forward_offset = None
return_val_offset1 = self.prev[self.prev[end]]
if (jump_back and jump_back != self.prev_op[end]
and self.is_jump_forward(jump_forward_offset)):
if (code[self.prev_op[end]] == self.opc.RETURN_VALUE or
(code[self.prev_op[end]] == self.opc.POP_BLOCK
and code[return_val_offset1] == self.opc.RETURN_VALUE)):
jump_back = None
if not jump_back:
jump_back = self.last_instr(start, end, self.opc.RETURN_VALUE)
if not jump_back:
return
jump_back += 2 # FIXME ???
if_offset = None
if code[self.prev_op[next_line_byte]] not in self.pop_jump_tf:
if_offset = self.prev[next_line_byte]
if if_offset:
loop_type = 'while'
self.ignore_if.add(if_offset)
else:
loop_type = 'for'
target = next_line_byte
end = xdis.next_offset(code[jump_back], self.opc, jump_back)
else:
if self.get_target(jump_back, 0) >= next_line_byte:
jump_back = self.last_instr(start, end, self.opc.JUMP_ABSOLUTE, start, False)
if end > jump_back+4 and self.is_jump_forward(end):
if self.is_jump_forward(jump_back+4):
if self.get_target(jump_back+4, extended_arg) == self.get_target(end, extended_arg):
self.fixed_jumps[offset] = jump_back+4
end = jump_back+4
elif target < offset:
self.fixed_jumps[offset] = jump_back+4
end = jump_back+4
# I think 0 right because jump_back has been adjusted for any EXTENDED_ARG
# it encounters
target = self.get_target(jump_back, 0)
if code[target] in (self.opc.FOR_ITER, self.opc.GET_ITER):
loop_type = 'for'
else:
loop_type = 'while'
test = self.prev_op[next_line_byte]
if test == offset:
loop_type = 'while 1'
elif self.code[test] in self.opc.JUMP_OPs:
self.ignore_if.add(test)
test_target = self.get_target(test, extended_arg)
if test_target > (jump_back+3):
jump_back = test_target
self.not_continue.add(jump_back)
self.loops.append(target)
self.structs.append({'type': loop_type + '-loop',
'start': target,
'end': jump_back})
after_jump_offset = xdis.next_offset(code[jump_back], self.opc, jump_back)
if after_jump_offset != end:
self.structs.append({'type': loop_type + '-else',
'start': after_jump_offset,
'end': end})
elif op in self.pop_jump_tf:
start = offset + instruction_size(op, self.opc)
target = self.get_target(offset, extended_arg)
rtarget = self.restrict_to_parent(target, parent)
prev_op = self.prev_op
# Do not let jump to go out of parent struct bounds
if target != rtarget and parent['type'] == 'and/or':
self.fixed_jumps[offset] = rtarget
return
# Does this jump to right after another conditional jump that is
# not myself? If so, it's part of a larger conditional.
# rocky: if we have a conditional jump to the next instruction, then
# possibly I am "skipping over" a "pass" or null statement.
if ((code[prev_op[target]] in self.pop_jump_if_pop) and
(target > offset) and prev_op[target] != offset):
# FIXME: this is not accurate The commented out below
# is what it should be. However grammar rules right now
# assume the incorrect offsets.
# self.fixed_jumps[offset] = target
self.fixed_jumps[offset] = prev_op[target]
self.structs.append({'type': 'and/or',
'start': start,
'end': prev_op[target]})
return
# The opcode *two* instructions before the target jump offset is important
# in making a determination of what we have. Save that.
pre_rtarget = prev_op[rtarget]
# Is it an "and" inside an "if" or "while" block
if op == self.opc.POP_JUMP_IF_FALSE and self.version < 3.6:
# Search for another POP_JUMP_IF_FALSE targetting the same op,
# in current statement, starting from current offset, and filter
# everything inside inner 'or' jumps and midline ifs
match = self.rem_or(start, self.next_stmt[offset],
self.opc.POP_JUMP_IF_FALSE, target)
# If we still have any offsets in set, start working on it
if match:
is_jump_forward = self.is_jump_forward(pre_rtarget)
if (is_jump_forward and pre_rtarget not in self.stmts and
self.restrict_to_parent(self.get_target(pre_rtarget, extended_arg), parent) == rtarget):
if (code[prev_op[pre_rtarget]] == self.opc.JUMP_ABSOLUTE
and self.remove_mid_line_ifs([offset]) and
target == self.get_target(prev_op[pre_rtarget], extended_arg) and
(prev_op[pre_rtarget] not in self.stmts or
self.get_target(prev_op[pre_rtarget], extended_arg) > prev_op[pre_rtarget]) and
1 == len(self.remove_mid_line_ifs(self.rem_or(start, prev_op[pre_rtarget], self.pop_jump_tf, target)))):
pass
elif (code[prev_op[pre_rtarget]] == self.opc.RETURN_VALUE
and self.remove_mid_line_ifs([offset]) and
1 == (len(set(self.remove_mid_line_ifs(self.rem_or(start, prev_op[pre_rtarget],
self.pop_jump_tf, target))) |
set(self.remove_mid_line_ifs(self.rem_or(start, prev_op[pre_rtarget],
(self.opc.POP_JUMP_IF_FALSE,
self.opc.POP_JUMP_IF_TRUE,
self.opc.JUMP_ABSOLUTE),
pre_rtarget, True)))))):
pass
else:
fix = None
jump_ifs = self.all_instr(start, self.next_stmt[offset],
self.opc.POP_JUMP_IF_FALSE)
last_jump_good = True
for j in jump_ifs:
if target == self.get_target(j, extended_arg):
if self.lines[j].next == j + 3 and last_jump_good:
fix = j
break
else:
last_jump_good = False
self.fixed_jumps[offset] = fix or match[-1]
return
else:
self.fixed_jumps[offset] = match[-1]
return
# op == POP_JUMP_IF_TRUE
else:
next = self.next_stmt[offset]
if prev_op[next] == offset:
pass
elif self.is_jump_forward(next) and target == self.get_target(next, extended_arg):
if code[prev_op[next]] == self.opc.POP_JUMP_IF_FALSE:
if (code[next] == self.opc.JUMP_FORWARD
or target != rtarget
or code[prev_op[pre_rtarget]] not in
(self.opc.JUMP_ABSOLUTE, self.opc.RETURN_VALUE)):
self.fixed_jumps[offset] = prev_op[next]
return
elif (code[next] == self.opc.JUMP_ABSOLUTE and self.is_jump_forward(target) and
self.get_target(target, extended_arg) == self.get_target(next, extended_arg)):
self.fixed_jumps[offset] = prev_op[next]
return
# Don't add a struct for a while test, it's already taken care of
if offset in self.ignore_if:
return
if (code[pre_rtarget] == self.opc.JUMP_ABSOLUTE and
pre_rtarget in self.stmts and
pre_rtarget != offset and
prev_op[pre_rtarget] != offset and
not (code[rtarget] == self.opc.JUMP_ABSOLUTE and
code[rtarget+3] == self.opc.POP_BLOCK and
code[prev_op[pre_rtarget]] != self.opc.JUMP_ABSOLUTE)):
rtarget = pre_rtarget
# Does the "jump if" jump beyond a jump op?
# That is, we have something like:
# POP_JUMP_IF_FALSE HERE
# ...
# JUMP_FORWARD
# HERE:
#
# If so, this can be block inside an "if" statement
# or a conditional assignment like:
# x = 1 if x else 2
#
# There are other contexts we may need to consider
# like whether the target is "END_FINALLY"
# or if the condition jump is to a forward location
if self.is_jump_forward(pre_rtarget):
if_end = self.get_target(pre_rtarget, 0)
# If the jump target is back, we are looping
if (if_end < pre_rtarget and
(code[prev_op[if_end]] == self.opc.SETUP_LOOP)):
if (if_end > start):
return
end = self.restrict_to_parent(if_end, parent)
self.structs.append({'type': 'if-then',
'start': start,
'end': pre_rtarget})
# FIXME: add this
# self.fixed_jumps[offset] = rtarget
self.not_continue.add(pre_rtarget)
if rtarget < end and (
code[rtarget] not in (self.opc.END_FINALLY,
self.opc.JUMP_ABSOLUTE) and
code[prev_op[pre_rtarget]] not in (self.opc.POP_EXCEPT,
self.opc.END_FINALLY)):
self.structs.append({'type': 'else',
'start': rtarget,
'end': end})
self.else_start[rtarget] = end
elif self.is_jump_back(pre_rtarget, 0):
if_end = rtarget
self.structs.append({'type': 'if-then',
'start': start,
'end': pre_rtarget})
self.not_continue.add(pre_rtarget)
elif code[pre_rtarget] in (self.opc.RETURN_VALUE,
self.opc.BREAK_LOOP):
self.structs.append({'type': 'if-then',
'start': start,
'end': rtarget})
# It is important to distingish if this return is inside some sort
# except block return
jump_prev = prev_op[offset]
if self.is_pypy and code[jump_prev] == self.opc.COMPARE_OP:
if self.opc.cmp_op[code[jump_prev+1]] == 'exception-match':
return
if self.version >= 3.5:
# Python 3.5 may remove as dead code a JUMP
# instruction after a RETURN_VALUE. So we check
# based on seeing SETUP_EXCEPT various places.
if code[rtarget] == self.opc.SETUP_EXCEPT:
return
# Check that next instruction after pops and jump is
# not from SETUP_EXCEPT
next_op = rtarget
if code[next_op] == self.opc.POP_BLOCK:
next_op += instruction_size(self.code[next_op], self.opc)
if code[next_op] == self.opc.JUMP_ABSOLUTE:
next_op += instruction_size(self.code[next_op], self.opc)
if next_op in targets:
for try_op in targets[next_op]:
come_from_op = code[try_op]
if come_from_op == self.opc.SETUP_EXCEPT:
return
pass
pass
if code[pre_rtarget] == self.opc.RETURN_VALUE:
self.return_end_ifs.add(pre_rtarget)
else:
self.fixed_jumps[offset] = rtarget
self.not_continue.add(pre_rtarget)
else:
# For now, we'll only tag forward jump.
if self.version >= 3.6:
if target > offset:
self.fixed_jumps[offset] = target
pass
else:
# FIXME: This is probably a bug in < 3.6 and we should
# instead use the above code. But until we smoke things
# out we'll stick with it.
if rtarget > offset:
self.fixed_jumps[offset] = rtarget
elif op == self.opc.SETUP_EXCEPT:
target = self.get_target(offset, extended_arg)
end = self.restrict_to_parent(target, parent)
self.fixed_jumps[offset] = end
elif op == self.opc.POP_EXCEPT:
next_offset = xdis.next_offset(op, self.opc, offset)
target = self.get_target(next_offset, extended_arg)
if target > next_offset:
next_op = code[next_offset]
if (self.opc.JUMP_ABSOLUTE == next_op and
self.opc.END_FINALLY != code[xdis.next_offset(next_op, self.opc, next_offset)]):
self.fixed_jumps[next_offset] = target
self.except_targets[target] = next_offset
elif op == self.opc.SETUP_FINALLY:
target = self.get_target(offset, extended_arg)
end = self.restrict_to_parent(target, parent)
self.fixed_jumps[offset] = end
elif op in self.jump_if_pop:
target = self.get_target(offset, extended_arg)
if target > offset:
unop_target = self.last_instr(offset, target, self.opc.JUMP_FORWARD, target)
if unop_target and code[unop_target+3] != self.opc.ROT_TWO:
self.fixed_jumps[offset] = unop_target
else:
self.fixed_jumps[offset] = self.restrict_to_parent(target, parent)
pass
pass
elif self.version >= 3.5:
# 3.5+ has Jump optimization which too often causes RETURN_VALUE to get
# misclassified as RETURN_END_IF. Handle that here.
# In RETURN_VALUE, JUMP_ABSOLUTE, RETURN_VALUE is never RETURN_END_IF
if op == self.opc.RETURN_VALUE:
next_offset = xdis.next_offset(op, self.opc, offset)
if (next_offset < len(code) and code[next_offset] == self.opc.JUMP_ABSOLUTE and
offset in self.return_end_ifs):
self.return_end_ifs.remove(offset)
pass
pass
elif op == self.opc.JUMP_FORWARD:
# If we have:
# JUMP_FORWARD x, [non-jump, insns], RETURN_VALUE, x:
# then RETURN_VALUE is not RETURN_END_IF
rtarget = self.get_target(offset, extended_arg)
rtarget_prev = self.prev[rtarget]
if (code[rtarget_prev] == self.opc.RETURN_VALUE and
rtarget_prev in self.return_end_ifs):
i = rtarget_prev
while i != offset:
if code[i] in [op3.JUMP_FORWARD, op3.JUMP_ABSOLUTE]:
return
i = self.prev[i]
self.return_end_ifs.remove(rtarget_prev)
pass
return
def build_statement_indices(self):
code = self.code
start = 0
end = codelen = len(code)
# Compose preliminary list of indices with statements,
# using plain statement opcodes
prelim = self.all_instr(start, end, self.statement_opcodes)
# Initialize final container with statements with
# preliminary data
stmts = self.stmts = set(prelim)
# Same for opcode sequences
pass_stmts = set()
for sequence in self.statement_opcode_sequences:
for i in self.op_range(start, end-(len(sequence)+1)):
match = True
for elem in sequence:
if elem != code[i]:
match = False
break
i += instruction_size(code[i], self.opc)
if match is True:
i = self.prev_op[i]
stmts.add(i)
pass_stmts.add(i)
# Initialize statement list with the full data we've gathered so far
if pass_stmts:
stmt_offset_list = list(stmts)
stmt_offset_list.sort()
else:
stmt_offset_list = prelim
# 'List-map' which contains offset of start of
# next statement, when op offset is passed as index
self.next_stmt = slist = []
last_stmt_offset = -1
i = 0
# Go through all statement offsets
for stmt_offset in stmt_offset_list:
# Process absolute jumps, but do not remove 'pass' statements
# from the set
if (code[stmt_offset] == self.opc.JUMP_ABSOLUTE
and stmt_offset not in pass_stmts):
# If absolute jump occurs in forward direction or it takes off from the
# same line as previous statement, this is not a statement
# FIXME: 0 isn't always correct
target = self.get_target(stmt_offset, 0)
if target > stmt_offset or self.lines[last_stmt_offset].l_no == self.lines[stmt_offset].l_no:
stmts.remove(stmt_offset)
continue
# Rewing ops till we encounter non-JUMP_ABSOLUTE one
j = self.prev_op[stmt_offset]
while code[j] == self.opc.JUMP_ABSOLUTE:
j = self.prev_op[j]
# If we got here, then it's list comprehension which
# is not a statement too
if code[j] == self.opc.LIST_APPEND:
stmts.remove(stmt_offset)
continue
# Exclude ROT_TWO + POP_TOP
elif (code[stmt_offset] == self.opc.POP_TOP
and code[self.prev_op[stmt_offset]] == self.opc.ROT_TWO):
stmts.remove(stmt_offset)
continue
# Exclude FOR_ITER + designators
elif code[stmt_offset] in self.designator_ops:
j = self.prev_op[stmt_offset]
while code[j] in self.designator_ops:
j = self.prev_op[j]
if code[j] == self.opc.FOR_ITER:
stmts.remove(stmt_offset)
continue
# Add to list another list with offset of current statement,
# equal to length of previous statement
slist += [stmt_offset] * (stmt_offset-i)
last_stmt_offset = stmt_offset
i = stmt_offset
# Finish filling the list for last statement
slist += [codelen] * (codelen-len(slist))
def detect_control_flow(self, offset, targets, inst_index):
"""
Detect type of block structures and their boundaries to fix optimized jumps
in python2.3+
"""
code = self.code
inst = self.insts[inst_index]
op = inst.opcode
# Detect parent structure
parent = self.structs[0]
start = parent["start"]
end = parent["end"]
# Pick inner-most parent for our offset
for struct in self.structs:
current_start = struct["start"]
current_end = struct["end"]
if (current_start <= offset < current_end) and (
current_start >= start and current_end <= end):
start = current_start
end = current_end
parent = struct
if self.version < 3.8 and op == self.opc.SETUP_LOOP:
# We categorize loop types: 'for', 'while', 'while 1' with
# possibly suffixes '-loop' and '-else'
# Try to find the jump_back instruction of the loop.
# It could be a return instruction.
start += inst.inst_size
target = self.get_target(offset)
end = self.restrict_to_parent(target, parent)
self.setup_loops[target] = offset
if target != end:
self.fixed_jumps[offset] = end
(line_no, next_line_byte) = self.lines[offset]
jump_back = self.last_instr(start, end, self.opc.JUMP_ABSOLUTE,
next_line_byte, False)
if jump_back:
jump_forward_offset = xdis.next_offset(code[jump_back],
self.opc, jump_back)
else:
jump_forward_offset = None
return_val_offset1 = self.prev[self.prev[end]]
if (jump_back and jump_back != self.prev_op[end]
and self.is_jump_forward(jump_forward_offset)):
if code[self.prev_op[end]] == self.opc.RETURN_VALUE or (
code[self.prev_op[end]] == self.opc.POP_BLOCK
and code[return_val_offset1] == self.opc.RETURN_VALUE):
jump_back = None
if not jump_back:
# loop suite ends in return
jump_back = self.last_instr(start, end, self.opc.RETURN_VALUE)
if not jump_back:
return
jb_inst = self.get_inst(jump_back)
jump_back = self.next_offset(jb_inst.opcode, jump_back)
if_offset = None
if code[self.prev_op[next_line_byte]] not in self.pop_jump_tf:
if_offset = self.prev[next_line_byte]
if if_offset:
loop_type = "while"
self.ignore_if.add(if_offset)
else:
loop_type = "for"
target = next_line_byte
end = xdis.next_offset(code[jump_back], self.opc, jump_back)
else:
if self.get_target(jump_back) >= next_line_byte:
jump_back = self.last_instr(start, end,
self.opc.JUMP_ABSOLUTE, start,
False)
jb_inst = self.get_inst(jump_back)
jb_next_offset = self.next_offset(jb_inst.opcode, jump_back)
if end > jb_next_offset and self.is_jump_forward(end):
if self.is_jump_forward(jb_next_offset):
if self.get_target(jb_next_offset) == self.get_target(
end):
self.fixed_jumps[offset] = jb_next_offset
end = jb_next_offset
elif target < offset:
self.fixed_jumps[offset] = jb_next_offset
end = jb_next_offset
target = self.get_target(jump_back)
if code[target] in (self.opc.FOR_ITER, self.opc.GET_ITER):
loop_type = "for"
else:
loop_type = "while"
test = self.prev_op[next_line_byte]
if test == offset:
loop_type = "while 1"
elif self.code[test] in self.opc.JUMP_OPs:
self.ignore_if.add(test)
test_target = self.get_target(test)
if test_target > (jump_back + 3):
jump_back = test_target
self.not_continue.add(jump_back)
self.loops.append(target)
self.structs.append({
"type": loop_type + "-loop",
"start": target,
"end": jump_back
})
after_jump_offset = xdis.next_offset(code[jump_back], self.opc,
jump_back)
if after_jump_offset != end:
self.structs.append({
"type": loop_type + "-else",
"start": after_jump_offset,
"end": end,
})
elif op in self.pop_jump_tf:
start = offset + inst.inst_size
target = inst.argval
rtarget = self.restrict_to_parent(target, parent)
prev_op = self.prev_op
# Do not let jump to go out of parent struct bounds
if target != rtarget and parent["type"] == "and/or":
self.fixed_jumps[offset] = rtarget
return
# Does this jump to right after another conditional jump that is
# not myself? If so, it's part of a larger conditional.
# rocky: if we have a conditional jump to the next instruction, then
# possibly I am "skipping over" a "pass" or null statement.
pretarget = self.get_inst(prev_op[target])
if (pretarget.opcode in self.pop_jump_if_pop and (target > offset)
and pretarget.offset != offset):
# FIXME: hack upon hack...
# In some cases the pretarget can be a jump to the next instruction
# and these aren't and/or's either. We limit to 3.5+ since we experienced there
# but it might be earlier versions, or might be a general principle.
if pretarget.argval != target:
# FIXME: this is not accurate The commented out below
# is what it should be. However grammar rules right now
# assume the incorrect offsets.
# self.fixed_jumps[offset] = target
self.fixed_jumps[offset] = pretarget.offset
self.structs.append({
"type": "and/or",
"start": start,
"end": pretarget.offset
})
return
# The opcode *two* instructions before the target jump offset is important
# in making a determination of what we have. Save that.
pre_rtarget = prev_op[rtarget]
if op == self.opc.POP_JUMP_IF_FALSE:
self.fixed_jumps[offset] = target
# op == POP_JUMP_IF_TRUE
else:
next = self.next_stmt[offset]
if prev_op[next] == offset:
pass
elif self.is_jump_forward(next) and target == self.get_target(
next):
if code[prev_op[next]] == self.opc.POP_JUMP_IF_FALSE:
if (code[next] == self.opc.JUMP_FORWARD
or target != rtarget
or code[prev_op[pre_rtarget]]
not in (self.opc.JUMP_ABSOLUTE,
self.opc.RETURN_VALUE)):
self.fixed_jumps[offset] = prev_op[next]
return
elif (code[next] == self.opc.JUMP_ABSOLUTE
and self.is_jump_forward(target)
and self.get_target(target) == self.get_target(next)):
self.fixed_jumps[offset] = prev_op[next]
return
rtarget_is_ja = code[pre_rtarget] == self.opc.JUMP_ABSOLUTE
if (rtarget_is_ja and pre_rtarget in self.stmts
and pre_rtarget != offset
and prev_op[pre_rtarget] != offset
and not (code[rtarget] == self.opc.JUMP_ABSOLUTE
and code[rtarget + 3] == self.opc.POP_BLOCK
and code[prev_op[pre_rtarget]] !=
self.opc.JUMP_ABSOLUTE)):
rtarget = pre_rtarget
# Does the "jump if" jump beyond a jump op?
# That is, we have something like:
# POP_JUMP_IF_FALSE HERE
# ...
# JUMP_FORWARD
# HERE:
#
# If so, this can be block inside an "if" statement
# or a conditional assignment like:
# x = 1 if x else 2
#
# For 3.5, for JUMP_FORWARD above we could have also
# JUMP_BACK or CONTINUE
#
# There are other situations we may need to consider, like
# if the condition jump is to a forward location.
# Also the existence of a jump to the instruction after "END_FINALLY"
# will distinguish "try/else" from "try".
rtarget_break = (self.opc.RETURN_VALUE, self.opc.BREAK_LOOP)
if self.is_jump_forward(pre_rtarget) or (rtarget_is_ja):
if_end = self.get_target(pre_rtarget)
# If the jump target is back, we are looping
if (if_end < pre_rtarget and self.version < 3.8
and (code[prev_op[if_end]] == self.opc.SETUP_LOOP)):
if if_end > start:
return
end = self.restrict_to_parent(if_end, parent)
self.structs.append({
"type": "if-then",
"start": start,
"end": pre_rtarget
})
# FIXME: add this
# self.fixed_jumps[offset] = rtarget
self.not_continue.add(pre_rtarget)
if rtarget < end and (
code[rtarget]
not in (self.opc.END_FINALLY, self.opc.JUMP_ABSOLUTE)
and code[prev_op[pre_rtarget]]
not in (self.opc.POP_EXCEPT, self.opc.END_FINALLY)):
self.structs.append({
"type": "else",
"start": rtarget,
"end": end
})
self.else_start[rtarget] = end
elif self.is_jump_back(pre_rtarget, 0):
if_end = rtarget
self.structs.append({
"type": "if-then",
"start": start,
"end": pre_rtarget
})
self.not_continue.add(pre_rtarget)
elif code[pre_rtarget] in rtarget_break:
self.structs.append({
"type": "if-then",
"start": start,
"end": rtarget
})
# It is important to distingish if this return is inside some sort
# except block return
jump_prev = prev_op[offset]
if self.is_pypy and code[jump_prev] == self.opc.COMPARE_OP:
if self.opc.cmp_op[code[jump_prev +
1]] == "exception-match":
return
pass
# Check that next instruction after pops and jump is
# not from SETUP_EXCEPT
next_op = rtarget
if code[next_op] == self.opc.POP_BLOCK:
next_op += instruction_size(self.code[next_op], self.opc)
if code[next_op] == self.opc.JUMP_ABSOLUTE:
next_op += instruction_size(self.code[next_op], self.opc)
if next_op in targets:
for try_op in targets[next_op]:
come_from_op = code[try_op]
if self.version < 3.8 and come_from_op == self.opc.SETUP_EXCEPT:
return
pass
self.fixed_jumps[offset] = rtarget
if code[pre_rtarget] == self.opc.RETURN_VALUE:
# If we are at some sort of POP_JUMP_IF and the instruction before was
# COMPARE_OP exception-match, then pre_rtarget is not an end_if
if not (inst_index > 0
and self.insts[inst_index - 1].argval
== "exception-match"):
self.return_end_ifs.add(pre_rtarget)
else:
self.fixed_jumps[offset] = rtarget
self.not_continue.add(pre_rtarget)
else:
if target > offset:
self.fixed_jumps[offset] = target
pass
elif self.version < 3.8 and op == self.opc.SETUP_EXCEPT:
target = self.get_target(offset)
end = self.restrict_to_parent(target, parent)
self.fixed_jumps[offset] = end
elif op == self.opc.POP_EXCEPT:
next_offset = xdis.next_offset(op, self.opc, offset)
target = self.get_target(next_offset)
if target > next_offset:
next_op = code[next_offset]
if (self.opc.JUMP_ABSOLUTE == next_op
and self.opc.END_FINALLY != code[xdis.next_offset(
next_op, self.opc, next_offset)]):
self.fixed_jumps[next_offset] = target
self.except_targets[target] = next_offset
elif op == self.opc.SETUP_FINALLY:
target = self.get_target(offset)
end = self.restrict_to_parent(target, parent)
self.fixed_jumps[offset] = end
elif op in self.jump_if_pop:
target = self.get_target(offset)
if target > offset:
unop_target = self.last_instr(offset, target,
self.opc.JUMP_FORWARD, target)
if unop_target and code[unop_target + 3] != self.opc.ROT_TWO:
self.fixed_jumps[offset] = unop_target
else:
self.fixed_jumps[offset] = self.restrict_to_parent(
target, parent)
pass
pass
else:
# 3.5+ has Jump optimization which too often causes RETURN_VALUE to get
# misclassified as RETURN_END_IF. Handle that here.
# In RETURN_VALUE, JUMP_ABSOLUTE, RETURN_VALUE is never RETURN_END_IF
if op == self.opc.RETURN_VALUE:
next_offset = xdis.next_offset(op, self.opc, offset)
if next_offset < len(code) and (
code[next_offset] == self.opc.JUMP_ABSOLUTE
and offset in self.return_end_ifs):
self.return_end_ifs.remove(offset)
pass
pass
elif op == self.opc.JUMP_FORWARD:
# If we have:
# JUMP_FORWARD x, [non-jump, insns], RETURN_VALUE, x:
# then RETURN_VALUE is not RETURN_END_IF
rtarget = self.get_target(offset)
rtarget_prev = self.prev[rtarget]
if (code[rtarget_prev] == self.opc.RETURN_VALUE
and rtarget_prev in self.return_end_ifs):
i = rtarget_prev
while i != offset:
if code[i] in [op3.JUMP_FORWARD, op3.JUMP_ABSOLUTE]:
return
i = self.prev[i]
self.return_end_ifs.remove(rtarget_prev)
pass
return
def detect_control_flow(self, offset, targets, inst_index):
"""
Detect type of block structures and their boundaries to fix optimized jumps
in python2.3+
"""
code = self.code
inst = self.insts[inst_index]
op = inst.opcode
# Detect parent structure
parent = self.structs[0]
start = parent['start']
end = parent['end']
# Pick inner-most parent for our offset
for struct in self.structs:
current_start = struct['start']
current_end = struct['end']
if ((current_start <= offset < current_end)
and (current_start >= start and current_end <= end)):
start = current_start
end = current_end
parent = struct
if self.version < 3.8 and op == self.opc.SETUP_LOOP:
# We categorize loop types: 'for', 'while', 'while 1' with
# possibly suffixes '-loop' and '-else'
# Try to find the jump_back instruction of the loop.
# It could be a return instruction.
start += inst.inst_size
target = self.get_target(offset)
end = self.restrict_to_parent(target, parent)
self.setup_loops[target] = offset
if target != end:
self.fixed_jumps[offset] = end
(line_no, next_line_byte) = self.lines[offset]
jump_back = self.last_instr(start, end, self.opc.JUMP_ABSOLUTE,
next_line_byte, False)
if jump_back:
jump_forward_offset = xdis.next_offset(code[jump_back], self.opc, jump_back)
else:
jump_forward_offset = None
return_val_offset1 = self.prev[self.prev[end]]
if (jump_back and jump_back != self.prev_op[end]
and self.is_jump_forward(jump_forward_offset)):
if (code[self.prev_op[end]] == self.opc.RETURN_VALUE or
(code[self.prev_op[end]] == self.opc.POP_BLOCK
and code[return_val_offset1] == self.opc.RETURN_VALUE)):
jump_back = None
if not jump_back:
# loop suite ends in return
jump_back = self.last_instr(start, end, self.opc.RETURN_VALUE)
if not jump_back:
return
jb_inst = self.get_inst(jump_back)
jump_back = self.next_offset(jb_inst.opcode, jump_back)
if_offset = None
if code[self.prev_op[next_line_byte]] not in self.pop_jump_tf:
if_offset = self.prev[next_line_byte]
if if_offset:
loop_type = 'while'
self.ignore_if.add(if_offset)
else:
loop_type = 'for'
target = next_line_byte
end = xdis.next_offset(code[jump_back], self.opc, jump_back)
else:
if self.get_target(jump_back) >= next_line_byte:
jump_back = self.last_instr(start, end, self.opc.JUMP_ABSOLUTE, start, False)
jb_inst = self.get_inst(jump_back)
jb_next_offset = self.next_offset(jb_inst.opcode, jump_back)
if end > jb_next_offset and self.is_jump_forward(end):
if self.is_jump_forward(jb_next_offset):
if self.get_target(jb_next_offset) == self.get_target(end):
self.fixed_jumps[offset] = jb_next_offset
end = jb_next_offset
elif target < offset:
self.fixed_jumps[offset] = jb_next_offset
end = jb_next_offset
target = self.get_target(jump_back)
if code[target] in (self.opc.FOR_ITER, self.opc.GET_ITER):
loop_type = 'for'
else:
loop_type = 'while'
test = self.prev_op[next_line_byte]
if test == offset:
loop_type = 'while 1'
elif self.code[test] in self.opc.JUMP_OPs:
self.ignore_if.add(test)
test_target = self.get_target(test)
if test_target > (jump_back+3):
jump_back = test_target
self.not_continue.add(jump_back)
self.loops.append(target)
self.structs.append({'type': loop_type + '-loop',
'start': target,
'end': jump_back})
after_jump_offset = xdis.next_offset(code[jump_back], self.opc, jump_back)
if after_jump_offset != end:
self.structs.append({'type': loop_type + '-else',
'start': after_jump_offset,
'end': end})
elif op in self.pop_jump_tf:
start = offset + inst.inst_size
target = inst.argval
rtarget = self.restrict_to_parent(target, parent)
prev_op = self.prev_op
# Do not let jump to go out of parent struct bounds
if target != rtarget and parent['type'] == 'and/or':
self.fixed_jumps[offset] = rtarget
return
# Does this jump to right after another conditional jump that is
# not myself? If so, it's part of a larger conditional.
# rocky: if we have a conditional jump to the next instruction, then
# possibly I am "skipping over" a "pass" or null statement.
pretarget = self.get_inst(prev_op[target])
if (pretarget.opcode in self.pop_jump_if_pop and
(target > offset) and pretarget.offset != offset):
# FIXME: hack upon hack...
# In some cases the pretarget can be a jump to the next instruction
# and these aren't and/or's either. We limit to 3.5+ since we experienced there
# but it might be earlier versions, or might be a general principle.
if self.version < 3.5 or pretarget.argval != target:
# FIXME: this is not accurate The commented out below
# is what it should be. However grammar rules right now
# assume the incorrect offsets.
# self.fixed_jumps[offset] = target
self.fixed_jumps[offset] = pretarget.offset
self.structs.append({'type': 'and/or',
'start': start,
'end': pretarget.offset})
return
# The opcode *two* instructions before the target jump offset is important
# in making a determination of what we have. Save that.
pre_rtarget = prev_op[rtarget]
# Is it an "and" inside an "if" or "while" block
if op == self.opc.POP_JUMP_IF_FALSE:
# Search for another POP_JUMP_IF_FALSE targetting the same op,
# in current statement, starting from current offset, and filter
# everything inside inner 'or' jumps and midline ifs
match = self.rem_or(start, self.next_stmt[offset],
self.opc.POP_JUMP_IF_FALSE, target)
# If we still have any offsets in set, start working on it
if match:
is_jump_forward = self.is_jump_forward(pre_rtarget)
if (is_jump_forward and pre_rtarget not in self.stmts and
self.restrict_to_parent(self.get_target(pre_rtarget), parent) == rtarget):
if (code[prev_op[pre_rtarget]] == self.opc.JUMP_ABSOLUTE
and self.remove_mid_line_ifs([offset]) and
target == self.get_target(prev_op[pre_rtarget]) and
(prev_op[pre_rtarget] not in self.stmts or
self.get_target(prev_op[pre_rtarget]) > prev_op[pre_rtarget]) and
1 == len(self.remove_mid_line_ifs(self.rem_or(start, prev_op[pre_rtarget], self.pop_jump_tf, target)))):
pass
elif (code[prev_op[pre_rtarget]] == self.opc.RETURN_VALUE
and self.remove_mid_line_ifs([offset]) and
1 == (len(set(self.remove_mid_line_ifs(self.rem_or(start, prev_op[pre_rtarget],
self.pop_jump_tf, target))) |
set(self.remove_mid_line_ifs(self.rem_or(start, prev_op[pre_rtarget],
(self.opc.POP_JUMP_IF_FALSE,
self.opc.POP_JUMP_IF_TRUE,
self.opc.JUMP_ABSOLUTE),
pre_rtarget, True)))))):
pass
else:
fix = None
jump_ifs = self.inst_matches(start, self.next_stmt[offset],
self.opc.POP_JUMP_IF_FALSE)
last_jump_good = True
for j in jump_ifs:
if target == self.get_target(j):
# FIXME: remove magic number
if self.lines[j].next == j + 3 and last_jump_good:
fix = j
break
else:
last_jump_good = False
self.fixed_jumps[offset] = fix or match[-1]
return
else:
if self.version < 3.6:
# FIXME: this is putting in COME_FROMs in the wrong place.
# Fix up grammar so we don't need to do this.
# See cf_for_iter use in parser36.py
self.fixed_jumps[offset] = match[-1]
elif target > offset:
# Right now we only add COME_FROMs in forward (not loop) jumps
self.fixed_jumps[offset] = target
return
# op == POP_JUMP_IF_TRUE
else:
next = self.next_stmt[offset]
if prev_op[next] == offset:
pass
elif self.is_jump_forward(next) and target == self.get_target(next):
if code[prev_op[next]] == self.opc.POP_JUMP_IF_FALSE:
if (code[next] == self.opc.JUMP_FORWARD
or target != rtarget
or code[prev_op[pre_rtarget]] not in
(self.opc.JUMP_ABSOLUTE, self.opc.RETURN_VALUE)):
self.fixed_jumps[offset] = prev_op[next]
return
elif (code[next] == self.opc.JUMP_ABSOLUTE and self.is_jump_forward(target) and
self.get_target(target) == self.get_target(next)):
self.fixed_jumps[offset] = prev_op[next]
return
# Don't add a struct for a while test, it's already taken care of
if offset in self.ignore_if:
return
rtarget_is_ja = code[pre_rtarget] == self.opc.JUMP_ABSOLUTE
if ( rtarget_is_ja and
pre_rtarget in self.stmts and
pre_rtarget != offset and
prev_op[pre_rtarget] != offset and
not (code[rtarget] == self.opc.JUMP_ABSOLUTE and
code[rtarget+3] == self.opc.POP_BLOCK and
code[prev_op[pre_rtarget]] != self.opc.JUMP_ABSOLUTE)):
rtarget = pre_rtarget
# Does the "jump if" jump beyond a jump op?
# That is, we have something like:
# POP_JUMP_IF_FALSE HERE
# ...
# JUMP_FORWARD
# HERE:
#
# If so, this can be block inside an "if" statement
# or a conditional assignment like:
# x = 1 if x else 2
#
# For 3.5, in addition the JUMP_FORWARD above we could have
# JUMP_BACK or CONTINUE
#
# There are other situations we may need to consider, like
# if the condition jump is to a forward location.
# Also the existence of a jump to the instruction after "END_FINALLY"
# will distinguish "try/else" from "try".
if self.version < 3.8:
rtarget_break = (self.opc.RETURN_VALUE, self.opc.BREAK_LOOP)
else:
rtarget_break = (self.opc.RETURN_VALUE,)
if self.is_jump_forward(pre_rtarget) or (rtarget_is_ja and self.version >= 3.5):
if_end = self.get_target(pre_rtarget)
# If the jump target is back, we are looping
if (if_end < pre_rtarget and self.version < 3.8 and
(code[prev_op[if_end]] == self.opc.SETUP_LOOP)):
if (if_end > start):
return
end = self.restrict_to_parent(if_end, parent)
self.structs.append({'type': 'if-then',
'start': start,
'end': pre_rtarget})
# FIXME: add this
# self.fixed_jumps[offset] = rtarget
self.not_continue.add(pre_rtarget)
if rtarget < end and (
code[rtarget] not in (self.opc.END_FINALLY,
self.opc.JUMP_ABSOLUTE) and
code[prev_op[pre_rtarget]] not in (self.opc.POP_EXCEPT,
self.opc.END_FINALLY)):
self.structs.append({'type': 'else',
'start': rtarget,
'end': end})
self.else_start[rtarget] = end
elif self.is_jump_back(pre_rtarget, 0):
if_end = rtarget
self.structs.append({'type': 'if-then',
'start': start,
'end': pre_rtarget})
self.not_continue.add(pre_rtarget)
elif code[pre_rtarget] in rtarget_break:
self.structs.append({'type': 'if-then',
'start': start,
'end': rtarget})
# It is important to distingish if this return is inside some sort
# except block return
jump_prev = prev_op[offset]
if self.is_pypy and code[jump_prev] == self.opc.COMPARE_OP:
if self.opc.cmp_op[code[jump_prev+1]] == 'exception-match':
return
if self.version >= 3.5:
# Python 3.5 may remove as dead code a JUMP
# instruction after a RETURN_VALUE. So we check
# based on seeing SETUP_EXCEPT various places.
if self.version < 3.6 and code[rtarget] == self.opc.SETUP_EXCEPT:
return
# Check that next instruction after pops and jump is
# not from SETUP_EXCEPT
next_op = rtarget
if code[next_op] == self.opc.POP_BLOCK:
next_op += instruction_size(self.code[next_op], self.opc)
if code[next_op] == self.opc.JUMP_ABSOLUTE:
next_op += instruction_size(self.code[next_op], self.opc)
if next_op in targets:
for try_op in targets[next_op]:
come_from_op = code[try_op]
if self.version < 3.8 and come_from_op == self.opc.SETUP_EXCEPT:
return
pass
pass
if self.version >= 3.4:
self.fixed_jumps[offset] = rtarget
if code[pre_rtarget] == self.opc.RETURN_VALUE:
# If we are at some sort of POP_JUMP_IF and the instruction before was
# COMPARE_OP exception-match, then pre_rtarget is not an end_if
if not (inst_index > 0 and self.insts[inst_index-1].argval == 'exception-match'):
self.return_end_ifs.add(pre_rtarget)
else:
self.fixed_jumps[offset] = rtarget
self.not_continue.add(pre_rtarget)
else:
# FIXME: this is very convoluted and based on rather hacky
# empirical evidence. It should go a way when
# we have better control-flow analysis
normal_jump = self.version >= 3.6
if self.version == 3.5:
j = self.offset2inst_index[target]
if j+2 < len(self.insts) and self.insts[j+2].is_jump_target:
normal_jump = self.insts[j+1].opname == 'POP_BLOCK'
if normal_jump:
# For now, we'll only tag forward jump.
if target > offset:
self.fixed_jumps[offset] = target
pass
else:
# FIXME: This is probably a bug in < 3.5 and we should
# instead use the above code. But until we smoke things
# out we'll stick with it.
if rtarget > offset:
self.fixed_jumps[offset] = rtarget
elif self.version < 3.8 and op == self.opc.SETUP_EXCEPT:
target = self.get_target(offset)
end = self.restrict_to_parent(target, parent)
self.fixed_jumps[offset] = end
elif op == self.opc.POP_EXCEPT:
next_offset = xdis.next_offset(op, self.opc, offset)
target = self.get_target(next_offset)
if target > next_offset:
next_op = code[next_offset]
if (self.opc.JUMP_ABSOLUTE == next_op and
self.opc.END_FINALLY != code[xdis.next_offset(next_op, self.opc, next_offset)]):
self.fixed_jumps[next_offset] = target
self.except_targets[target] = next_offset
elif op == self.opc.SETUP_FINALLY:
target = self.get_target(offset)
end = self.restrict_to_parent(target, parent)
self.fixed_jumps[offset] = end
elif op in self.jump_if_pop:
target = self.get_target(offset)
if target > offset:
unop_target = self.last_instr(offset, target, self.opc.JUMP_FORWARD, target)
if unop_target and code[unop_target+3] != self.opc.ROT_TWO:
self.fixed_jumps[offset] = unop_target
else:
self.fixed_jumps[offset] = self.restrict_to_parent(target, parent)
pass
pass
elif self.version >= 3.5:
# 3.5+ has Jump optimization which too often causes RETURN_VALUE to get
# misclassified as RETURN_END_IF. Handle that here.
# In RETURN_VALUE, JUMP_ABSOLUTE, RETURN_VALUE is never RETURN_END_IF
if op == self.opc.RETURN_VALUE:
next_offset = xdis.next_offset(op, self.opc, offset)
if ( next_offset < len(code) and
(code[next_offset] == self.opc.JUMP_ABSOLUTE and
offset in self.return_end_ifs) ):
self.return_end_ifs.remove(offset)
pass
pass
elif op == self.opc.JUMP_FORWARD:
# If we have:
# JUMP_FORWARD x, [non-jump, insns], RETURN_VALUE, x:
# then RETURN_VALUE is not RETURN_END_IF
rtarget = self.get_target(offset)
rtarget_prev = self.prev[rtarget]
if (code[rtarget_prev] == self.opc.RETURN_VALUE and
rtarget_prev in self.return_end_ifs):
i = rtarget_prev
while i != offset:
if code[i] in [op3.JUMP_FORWARD, op3.JUMP_ABSOLUTE]:
return
i = self.prev[i]
self.return_end_ifs.remove(rtarget_prev)
pass
return
